Research On Feature Detection Of Wireless Charging Spot And Vehicle Pose Estimation Based On Deep Learning

In recent years,with the rapid development of electric vehicles and wireless charging technology,it is an inevitable trend to use wireless charging technology to recharge electric vehicles in the feature.However,wireless charging efficiency is relatively low,especially when there is a misalignment between transmitter coil and receiver coil,which is an important factor restricting the development of wireless charging for electric vehicles.This is mainly reflected in the existence of blind area in vehicle positioning,and the driver can't observe whether transmitter coil and receiver coil on the vehicle chassis are accurately aligned or not;In addition,due to the complex parking environment,the traditional machine vision feature detection algorithms have the disadvantages of slow detection speed,insufficient robustness,high requirements for the environment and fuzzy detection results.According to the above problems,this paper refers on the successful experience of the application of machine learning or deep learning in the fields of classification,recognition and target detection,and then proposes a method based on deep learning for vehicle parking feature detection and vehicle pose estimation,which provides an important reference for realizing accurate alignment of wireless charging vehicles.The main research contents and achievements include:1.Research the feature corner detection algorithm of wireless charging spot based on deep learning in various extreme scenarios.The images of the parking spot data were collected and labeled with data samples,and based on the convolutional neural network designed by yolov3-tiny,the vehicle parking spot features were trained and tested,getting the class information and pixel position information of each feature corner.Besides,the correctness of detecting feature point categories and image coordinate information is proved by experiments.2.Aiming at the condition that there are only a few feature points in the actual application scene,the camera pose estimation method of sparse points is studied,and the camera pose estimation under the condition of 3 and 4 feature points is realized by using deep learning and P4 P method.Based on the comparative analysis results of the above sparse point camera pose estimation errors,and considering the actual situation of the precise positioning application scenario of wireless charging vehicles,a two-step vehicle pose estimation scheme for this application is proposed.In other words,when the vehicle is unable to detect the information of four feature points,the 3-point sparse pose estimation based on deep learning is adopted first.Once the number of feature points detected by the system meets the condition of P4 P method,the P4 P method is adopted to effectively estimate the vehicle pose with insufficient feature points.3.An experimental test system was built.After detecting the parking image collected by the camera through the parking spot feature points,the class information and pixel position information of each feature corner point of the parking spot were output.The information was input the vehicle pose estimation system,and then the camera pose information was output.The experimental results show that the proposed scheme can realize the estimation of vehicle pose,and prove the feasibility and correctness of the prediction of vehicle position and the estimation of vehicle pose using deep learning in the multi-extreme environment.